A $250 Nvidia AI board ran a BAE Systems strike drone end-to-end, no radio link

A single Nvidia Jetson Orin Nano board, the $250 palm-sized AI computer more often found in robots and smart cameras, ran the entire find, rank, and strike loop of a BAE Systems Bofors loitering munition at Winter Demo 2026 in Sweden. No radio link. About 320 seconds total. The mission completed at roughly -18°C.

Scaleout Systems, the Swedish AI startup that built the onboard software, used a YOLOv8 Nano object-detection model on the small Nvidia board, not a frontier model, to detect and rank targets in roughly 200 seconds of search, then fly to and release an explosive on the top-ranked vehicle, an armored engineering target, per Tom's Hardware and Ars Technica. The mission ran fully onboard, with no external communications, a deliberate choice for resilience against electronic warfare rather than a failure mode. The pilot was reduced to a single button press to start and remained a failsafe controller throughout.

The airframe was BAE Systems Bofors' Affordable Loitering Modular Ammunition, or ALMA, an existing one-way attack drone that circles a target before striking, flown in its Airolit S1 configuration according to the Nordic Defence Sector release on the demo. The loitering-munition framing matters here: media coverage has called it a kamikaze drone, but the program is a one-way attack drone the human operator sets loose within engagement parameters, not a free-hunting autonomous weapon. The pilot's start button is the human-in-the-loop gate, and the failsafe controller is the off-ramp.

The targeting loop, which historically depended on a large model, a GPU cluster, or a high-bandwidth satellite link, fit on a single low-power board at Winter Demo 2026. The drone ran at roughly 30 frames per second while moving at about 20 meters per second, with target latency in the field reported at around 30 milliseconds per Scaleout's demo write-up. Range estimation worked without a depth sensor by combining a pinhole camera model with a known object size for the target. The Nvidia board is the same kind of device used in robotics and edge cameras, and it is sold commercially for roughly $250.

The deliberate removal of the radio link changes the policy stakes. A drone that cannot be jammed because it never had a data link is a different operational problem than a drone that loses its link in a contest. Scaleout positions this as a resilience feature, and the demo was designed to prove the claim. The mission completed in comms-denied conditions, and the same board handled target search, ranking, navigation, and terminal strike without a fallback channel.

Scaleout is a startup with a single controlled demo at a Swedish air force range, not a combat record. The "Tactical Computer Vision Network" the company describes, which would let drones train locally and share model updates across a federated learning network, appears in company materials and is unproven in the field. The parent project, FEDAIR, sits inside NATO's DIANA accelerator, a NATO program that funds dual-use defense startups, with reported development funding of about 100,000 euros (roughly $115,000) plus training and test access. Per the University of Uppsala release on the program, the funding is small by defense-acquisition standards, and the technical claims scale with future iterations, not this one demo.

DroneXL's coverage of the demo notes the integration is the first public showing, and the open question is whether EW-resilient autonomy on a single chip moves from a controlled demo to a tested operational concept, and on what timeline. Scaleout's demo video is titled "Technical Demo: Onboard Edge Intelligence for Autonomous UAV Missions," and the company has not announced a follow-up date. BAE Systems Bofors has not announced a production timeline for ALMA with the Scaleout stack.